To address the limitations of traditional outlier detection methods for seepage monitoring data in processing nonlinear and nonstationary water level monitoring data, a hybrid model based on variational mode decomposition (VMD), long short-term memory network (LSTM), and variational autoencoder (VAE) is proposed. In the proposed model, VMD is used to perform multiscale decomposition of the water level time series and obtain intrinsic mode functions with different frequency characteristics. A two-layer LSTM is then employed to capture the temporal dependencies of each modal component. VAE is used for feature compression and reconstruction, and outlier detection is achieved based on the reconstruction error. Case validation results show that the hybrid model can effectively identify isolated outliers and continuous outlier sequences. In scenarios where outliers were randomly added to the test set at proportions of 1%, 2%, and 3%, the model exhibited good detection performance, verifying its effectiveness in outlier detection for seepage monitoring data.